Efficient mapping of ADAS (Advanced Driver Assistance System) algorithms onto multicore architectures

This thesis aims to address the real-time performance requirements of ADAS(Advanced Driver Assistance System) and autonomous vehicle applications on emerging multicore CPU and manycore GPU architectures. A parallel particle filter basedvehicle localization and map matching algorithm which fuses GPS, odometer and digital maps, and a parallel template-matching based traffic sign recognition algorithm which employes a Kinect sensor and digital map fusion are proposed. Implementations were performed on multicore CPUs using OpenMP programming model and on manycore GPUs using CUDA programming model. Real data were collected via a vehicle equipped with sensors for various road and weather conditions and performance tests were conducted on a parallel system having two six-core CPUs and two 512-cores GPUs. The execution times and speedup of parallel processing is examined. The effect of number of particles on the success rate of the localization algorithm is also observed. Test results show that up to 75 times speedups for particle filter based localization and mapmatching algorithm and up to 35 times speedups for the traffic sign recognition algorithm can be achieved on GPUs compared to implementations on sequential systems, and evidently the algorithms can be used with real-time performance in the vehicle environment. It is concluded that the emerging general purpose multicore/manycore processors can constitute a unified vehicle computing platform where ADAS applications can be implemented in parallel and run with real-time performances by replacing specialized hardware and/or software platforms used for each application.